pm.MCMC in pymc3 or in an other library?

Questions
1

Hi,
I am currently using the pymc library with python 3.6 to use the very useful function pm.MCMC.
I have not found its equivalent in pymc3, does the library exist in pymc3?

Many thanks in advance
Nicolas

2

You’re probably looking for pm.sample(). Note that things work quite differently in pymc3, but all sampling is done by calls to this and related these methods (described at that link).

1 Like
3

Thank you Christian,

I am currently looking for transfering a pymc2 code to pymc3 or another library compatible with python 3.8
In pymc2, I have :
U_pymc_model = pm.stochastic(logp=U_law_Logp, random=U_law_rand,
doc=‘Vit’, name=‘Vit’, parents={}, trace=True)

and
sampler = pm.MCMC([U_pymc_model, Hs_pymc_model, Tp_pymc_model, WdDr_pymc_model, WvDr_pymc_model,sigU_pymc_model, metamodel_response, refinement_criterion])

And I don’t know how to make them compatible with python 3.8
Many thanks in advance
Nicolas

4

Someone who is more well-versed in pymc2 should weigh in (@twiecki maybe? @junpenglao ?), but the pymc3 snippet below might get you in the vicinity. I will also mention that a “straight” pymc3 port of your pymc2 code may not be the most efficient/natural way to go. But if you have loads of pymc2 code, it might make sense.

Note that pymc3 will automatically a select a sampler based on your model, so you aren’t required to specify this before calling pm.sample() (though you can if you wish).

with pm.Model() as my_model:
    Vit = pm.DensityDist(
        'Vit',
        U_law_Logp,
        random=U_law_rand
    )
    trace = pm.sample()
5

Hi everybody,
Thank you Christian for your answer. I But have an issue with the solution you suggested.
1- I have also trouble installing pymc3. Which version should I install easely and without dependancy with python 3.8 ?
2- I would like to rewrite these lines :
U_pymc_model = pm.stochastic(logp=U_law_Logp, random=U_law_rand, doc=‘Vit’, name=‘Vit’, parents={}, trace=True)
sampler = pm.MCMC([U_pymc_model, metamodel_response, refinement_criterion])
in a python version 3.8, is it possible ? And if yes, how can I do it ?
Many thanks in advance
Nicolas

6

Installation instructions can be found on the main documentation page. Dependencies can be found here.

Regarding a rewrite, I suspect that a pymc3 version will not simply replace those 2 old lines of code with 2 new lines of code. As I said, pymc3 does not work in the same way that pymc2 does and so a line-by-line translation may not be possible.

1 Like
7

Hi Nicolas,

I think you should abandon PyMC2 completely and embrace PyMC3. Besides using up-to-date samplers, the code base of PyMC3 is just so much better. It’s a joy to work with compared to PyMC2.

Unfortunately though, this does mean throwing out your old code.

8

Thanks you Sammosummo, but it is not simple for me to to transform a pm.stochastic object into a pymc3 object. Is it possible to habve some help?
Many thanks in avance
Nicolas

9

Could you post a complete minimal example of your PyMC2 code, with a subset of or simulated data?

1 Like
10

Hi, thank you very much for your reply. Below a minimal example of a code that I want to convert into pymc3.
Many thanks in advance
Nicolas
KA_U_TI_ref_pour_portage_pymc3.py (2.5 KB)

11

Hi again, I can add a more simplified example if necessary. Just let me know.
Regards
Nicolas